Meta’s open-weight return now has a flagship, and it does not carry the Llama name. Muse Spark Open landed on 15 October with 816B parameters, a 1M token context and a license that makes every product built on it say so, two months after the 30B Muse Glimmer moved Meta’s open models to the Muse brand. October also brought DeepSeek V4.1-Flash, a new 552B base rather than a re-tune of V4 Flash, and two small MoE models you can run on one GPU: LFM2.5-24B-A2B and MiniCPM-V 4.7.
This open source LLM comparison covers every major open-weight family: benchmarks, license gates, download sizes, and what each model needs under Ollama or llama.cpp, where unified memory and VRAM decide what fits. Parameter counts, licenses and context lengths come from the model cards, config.json files and LICENSE files on Hugging Face, with the October releases read in October 2026. The hands-on numbers are our own runs: small models on a 4 vCPU, 16 GB Ubuntu 24.04 VM in March 2026, and DeepSeek V4 Flash and Kimi K3 in August.
Master Comparison Table
This table covers every major open-source/open-weight LLM family. “Active params” refers to how many parameters are used per inference pass in Mixture-of-Experts (MoE) architectures. Dense models use all parameters on every token.
| Model | Developer | Total Params | Active Params | Architecture | Context Window | Multimodal | License | Release |
|---|---|---|---|---|---|---|---|---|
| Kimi K3 | Moonshot AI | 2.8T | 104B | MoE (896 experts) | 1M tokens | Text + Image + Video | Kimi K3 (custom) | Jul 2026 |
| Kimi K2 | Moonshot AI | 1T | 32B | MoE (384 experts) | 128K (256K on K2 Thinking) | No | Modified MIT | Jul 2025 |
| DeepSeek V4.1-Flash (1017) | DeepSeek | 552B backbone + 196B Engram memory | 8B prefill / 16B decode | Causal encoder-decoder MoE (384 experts + 1 shared, 6 active) + vision encoder | 1M tokens | Text + Image | MIT | Oct 2026 |
| DeepSeek V4 Pro (0813) | DeepSeek | 1.65T | 49B | MoE (384 experts + 1 shared, 6 active) + sparse attention | 1M tokens | No | MIT | Aug 2026 |
| DeepSeek V4 Flash Vision (Exp) | DeepSeek | 305B | 13B + vision encoder | MoE (256 experts + 1 shared) + vision encoder | 1M tokens | Text + Image | MIT | Aug 2026 |
| DeepSeek V4 Pro (Preview) | DeepSeek | 1.6T | 49B | MoE + sparse attention | 1M tokens | No | MIT | Apr 2026 |
| DeepSeek V4 Flash | DeepSeek | 284B | 13B | MoE + sparse attention | 1M tokens | No | MIT | Apr 2026 |
| DeepSeek V3.2 | DeepSeek | 671B | 37B | MoE | 128K tokens | No | MIT | Dec 2025 |
| DeepSeek R1 | DeepSeek | 671B | 37B | MoE | 128K tokens | No | MIT | Jan 2025 |
| GLM-5.3 | Z.ai (Zhipu) | 753B | ~40B | MoE (256 experts + 1 shared, 8 active) | 1M tokens | No | GLM-5.3 (custom) | Aug 2026 |
| GLM-5.3-Flash | Z.ai (Zhipu) | 320B | 18B | MoE (288 experts + 1 shared, 8 active), hybrid sparse + linear attention | 1M tokens | Text + Image | MIT | Aug 2026 |
| GLM-5.2 | Z.ai (Zhipu) | 753B | ~40B | MoE | 1M tokens | No | MIT | Jun 2026 |
| GLM-5 | Z.ai (Zhipu) | 744B | 40B | MoE | 205K tokens | No | MIT | Feb 2026 |
| Qwen3.8-Max (2.4T-A95B) | Alibaba | 2.45T | 95B | MoE (512 experts, 10 routed + 1 shared) | 262K native (1M extensible) | No (open weights are text-only) | Qwen3.8-Max (custom) | Aug 2026 |
| Qwen3.8-Flash-Next | Alibaba | 180B on disk (125B plus 51B n-gram embeddings and 4B MTP) | 6B | MoE (512 experts, 10 routed + 1 shared) | 262K native (1M extensible) | Text + Image + Video | Qwen Community 1.0 | Aug 2026 |
| Qwen3.8-27B | Alibaba | 27.8B | 27.8B | Dense (native vision-language) | 262K native (1M extensible) | Text + Image + Video | Apache 2.0 | Aug 2026 |
| Qwen3.6-27B | Alibaba | 27B | 27B | Dense (hybrid attention) | 262K (1M via YaRN) | Text + Image + Video | Apache 2.0 | Apr 2026 |
| Qwen3.6-35B-A3B | Alibaba | 35B | 3B | MoE (hybrid attention) | 262K (1M via YaRN) | Text + Image + Video | Apache 2.0 | Apr 2026 |
| Qwen 3.5 397B-A17B | Alibaba | 397B | 17B | MoE | 256K tokens | Text + Image | Apache 2.0 | Feb 2026 |
| Qwen 3.5 122B-A10B | Alibaba | 122B | 10B | MoE | 256K tokens | Text + Image | Apache 2.0 | Feb 2026 |
| Qwen 3.5 27B | Alibaba | 27B | 27B | Dense | 256K tokens | Text + Image | Apache 2.0 | Feb 2026 |
| Qwen 3 235B | Alibaba | 235B | 22B | MoE (128e, 8 active) | 128K tokens | No | Apache 2.0 | Apr 2025 |
| Qwen 3 8B | Alibaba | 8B | 8B | Dense | 128K tokens | No | Apache 2.0 | Apr 2025 |
| gpt-oss-120b | OpenAI | 117B | 5.1B | MoE (native MXFP4) | 128K tokens | No | Apache 2.0 | Aug 2025 |
| gpt-oss-20b | OpenAI | 21B | 3.6B | MoE (native MXFP4) | 128K tokens | No | Apache 2.0 | Aug 2025 |
| Muse Spark Open | Meta | 816B | 42B | MoE (256 experts, 8 active) + 1.8B perception encoder | 1M tokens | Text + Image + Video + Audio | Muse Model License | Oct 2026 |
| Muse Glimmer 30B | Meta | 29.6B | 29.6B | Dense + 1.8B perception encoder | 131K tokens | Text + Image | Apache 2.0 | Aug 2026 |
| Llama 4 Scout | Meta | 109B | 17B | MoE (16 experts) | 10M tokens | Text + Image | Llama 4 Community | Apr 2025 |
| Llama 4 Maverick | Meta | 400B | 17B | MoE (128 experts) | 1M tokens | Text + Image | Llama 4 Community | Apr 2025 |
| Llama 3.3 | Meta | 70B | 70B | Dense | 128K tokens | No | Llama 3.3 Community | Dec 2024 |
| Gemma 4 31B | 30.7B | 30.7B | Dense | 256K tokens | Text + Image | Apache 2.0 | Mar 2026 | |
| Gemma 4 26B A4B | 25.2B | 3.8B | MoE (128e, 8 active) | 256K tokens | Text + Image | Apache 2.0 | Mar 2026 | |
| Gemma 4 12B | 12B | 12B | Dense (encoder-free) | 256K tokens | Text + Image + Audio | Apache 2.0 | Jun 2026 | |
| Gemma 4 E4B | 8B | 4.5B effective | Dense (per-layer embeddings) | 128K tokens | Text + Image + Audio | Apache 2.0 | Mar 2026 | |
| Gemma 3 27B | 27B | 27B | Dense | 128K tokens | Text + Image | Gemma (agreement) | Mar 2025 | |
| Mistral Small 4 | Mistral AI | 119B | 6B | MoE (128e, 4 active) | 256K tokens | Text + Image | Apache 2.0 | Mar 2026 |
| Mistral Large 3 | Mistral AI | 675B | 41B | MoE | 256K tokens | Text + Image | Apache 2.0 | Dec 2025 |
| Phi-4 Reasoning Vision | Microsoft | 15B | 15B | Dense | 16K tokens | Text + Image | MIT | Mar 2026 |
| Phi-4 | Microsoft | 14B | 14B | Dense | 16K tokens | No | MIT | Jan 2025 |
| Phi-4 Mini | Microsoft | 3.8B | 3.8B | Dense | 128K tokens | No | MIT | Jan 2025 |
| MiniCPM-V 4.7 35B-A3B | OpenBMB | 35.2B | 3B | MoE (256 experts, 8 active) + vision encoder | 262K (config) | Text + Image + Video | Apache 2.0 | Oct 2026 |
| LFM2.5-24B-A2B | Liquid AI | 24B | 2.45B | Hybrid MoE (30 short-convolution + 10 attention layers, 64 experts, 4 active) | 131K tokens | No | LFM Open License v1.0 | Oct 2026 |
| Command A | Cohere | 111B | 111B | Dense | 256K tokens | No | CC-BY-NC | Mar 2025 |
| Falcon 3 10B | TII Abu Dhabi | 10B | 10B | Dense | 32K tokens | No | TII Falcon-LLM 2.0 | Dec 2024 |
Benchmark Scores
Benchmarks tell part of the story. MMLU-Pro tests professional-level knowledge, GPQA Diamond tests graduate-level reasoning, and SWE-bench Verified measures real-world coding ability against actual GitHub issues. All scores below come from official model cards or the developers’ published evaluations. Empty cells mean the score has not been published or independently verified. Watch the benchmark names carefully: SWE-bench Pro (used by Z.ai for GLM-5.2) is a harder, newer suite than SWE-bench Verified, so those numbers are not comparable across columns.
| Model | MMLU-Pro | GPQA Diamond | SWE-bench Verified | Other verified highlights |
|---|---|---|---|---|
| DeepSeek V4.1-Flash | N/A | 90.9% (max effort) | N/A | Terminal-Bench 2.1 90.6, DeepSWE v1.1 74.2; the 1017 update reports DeepSWE 77.0 |
| Muse Spark Open | N/A | 89.0% (xhigh) | N/A | MCP-Atlas 89.2, OSWorld Verified 80.4, Terminal-Bench 2.1 76.4 (Terminus 2 agent) |
| Muse Glimmer 30B | N/A | 83.5% (AA run) | 76.0% (high reasoning) | SWE-bench Pro 51.2, MCP Atlas 75.5 |
| DeepSeek V4 Pro | N/A | N/A | 80.6% (max thinking) | Highest published open-weight SWE-bench Verified score |
| GLM-5 | N/A | N/A | 77.8% | Humanity’s Last Exam 50.4 |
| Qwen3.6-27B | N/A | N/A | 77.2% | Beats Qwen 3.5 397B MoE (76.2%) on the same suite |
| Qwen3.6-35B-A3B | 85.2% | 86.0% | 73.4% | 3B active params |
| gpt-oss-120b | N/A | 80.1% (high reasoning) | 62.4% (high reasoning) | 52.6% SWE-bench Verified at medium reasoning effort |
| GLM-5.2 | N/A | 91.2% | N/A | SWE-bench Pro 62.1, AIME 2026 99.2, Terminal-Bench 2.1 81.0 |
| DeepSeek R1 | 84.0% | 71.5% | N/A | MATH-500 97.3% |
| Qwen 3 235B | 83.8% | 77.1% | N/A | AIME ’24 85.7% |
A few things jump out. DeepSeek V4 Pro’s 80.6% on SWE-bench Verified is the strongest published open-weight result on that benchmark, and the smaller V4 Flash inherits the same sparse-attention architecture at a fraction of the memory footprint. Qwen3.6-27B posting 77.2% is arguably the more remarkable number: that is a dense 27B model you can run on a single 24 GB consumer GPU, beating Alibaba’s own 397B MoE flagship from two months earlier. On the Artificial Analysis Intelligence Index, checked on 22 October 2026, DeepSeek V4.1-Flash 1017 at maximum effort ties GLM-5.3 at maximum on 45, the highest open-weight score on this page (the hosted Qwen3.8 Max also lists 45), ahead of Kimi K3 on 44, GLM-5.3-Flash on 42, Qwen3.8-2.4T-A95B and Qwen3.8-Flash-Next on 40, Muse Spark Open on 39, and DeepSeek V4 Pro 0813 on 36. Meta’s own Muse card cites the same 39. DeepSeek’s V4.1-Flash card also posts 90.6 on Terminal-Bench 2.1, the highest open-weight figure on this page, though it was run in DeepSeek’s own minimal harness. Leaderboard positions move, so look it up rather than trusting any snapshot, including this one.
Open flagships against closed models on one comparison set
Cross-model numbers are usually impossible to line up because every vendor picks a different suite. Z.ai’s GLM-5.3 card is the exception: it reports one set of benchmarks across five open-weight flagships and three closed models, run under stated harnesses. The table below is a ten-row, six-column subset of the sixteen rows on that card, dropping the GLM-5.2 and Fable 5 columns for width. Treat it as Z.ai-reported rather than independent, because Z.ai chose which benchmarks appear.
| Benchmark | GLM-5.3 | Kimi K3 | DeepSeek V4 Pro (0813) | Qwen3.8-Max | Opus 4.8 | GPT-5.6 Sol |
|---|---|---|---|---|---|---|
| Terminal Bench 2.1 | 88.2 | 88.3 | 87.9 | 86.6 | 85.0 | 88.8 |
| Terminal Bench 3.0 | 28.3 | 17.4 | n/r | n/r | 21.1 | 34.6 |
| DeepSWE (v1.1) | 66.9 | 67.5 | 62.7 | 56.6 | 58.0 | 72.7 |
| NL2Repo | 58.0 | 58.0 | 61.1 | 55.9 | 69.7 | n/r |
| CyberGym | 84.5 | 80.0 | 83.3 | 78.5 | 78.1 | 83.6 |
| Toolathlon Verified | 73.0 | 76.5 | 74.1 | 72.5 | 76.2 | 74.9 |
| AutomationBench (v1.0.6) | 48.2 | 46.7 | 43.2 | 39.8 | 41.0 | 45.8 |
| Agents’ Last Exam (CLI) | 28.5 | 27.6 | 25.7 | 27.0 | 25.7 | 28.6 |
| HLE with tools | 62.5 | 59.8 | 60.0 | 56.2 | 57.9 | 64.5 |
| GDPval-AA v2 | 1769 | 1682 | 1590 | 1739 | 1588 | 1730 |
Counting outright wins across those ten rows: GPT-5.6 Sol takes five, GLM-5.3 three, and Kimi K3 and Opus 4.8 one each. The best open-weight score leads on four of the ten. Where the open models win they mostly win narrowly: 0.3 on Toolathlon, 0.7 on CyberGym, and 2.0 on AutomationBench, each measured against the best closed-model score on that row, including the Fable 5 column dropped above, with GDPval-AA v2 the one clear margin. Where they lose they lose by more: 8.6 points behind Opus 4.8 on repo-level code generation, 6.3 behind GPT-5.6 Sol on the newest terminal benchmark, and 5.2 on DeepSWE. The gap that remains is in long-horizon code generation, not in tool use or security work, where open weights are now at parity.
GLM-5.3 improved on GLM-5.2 in every one of the sixteen rows on its card, and the two largest gains among the ten shown here are on benchmarks GLM-5.2 barely registered: Terminal Bench 3.0 went from 4.6 to 28.3 and AutomationBench from 26.2 to 48.2. That is a post-training result on an unchanged base model, which is the part worth internalizing about this release cycle.
Do vendors agree on each other’s scores?
Mostly yes, which is not what you would expect from marketing material. DeepSeek V4 Pro appears on both Z.ai’s card and DeepSeek’s own, so its rows can be checked against each other. Of the eight benchmarks both vendors report for it, six are numerically identical: Terminal Bench 2.1 at 87.9, DeepSWE at 62.7, CyberGym at 83.3, Toolathlon at 74.1, Agents’ Last Exam at 25.7, and HLE with tools at 60.0. NL2Repo differs by 0.4, with Z.ai reporting 61.1 against DeepSeek’s 61.5.
The eighth row is the useful warning. Both cards list an AutomationBench score for the same model, 43.2 on one and 31.8 on the other, an 11.4 point spread that looks like a contradiction until you read the labels: Z.ai ran version 1.0.6 and DeepSeek ran the public split. Different benchmark, same name. Widening the check to Kimi K3, Opus 4.8, and GLM-5.2, 17 of their 19 shared rows match to the decimal. Two do not, and neither carries a version suffix to explain it: Kimi K3’s HLE with tools reads 59.8 on Z.ai’s card against 56.0 on DeepSeek’s, a 3.8 point spread, and Opus 4.8’s CyberGym differs by 0.2. Fable 5 sits on both cards too and is looser than any of them, with one of its five comparable rows matching and Toolathlon the widest at 74.7 against 77.9. So vendor-reported numbers travel between cards far better than you would expect, though not perfectly, and two cards agreeing on a benchmark name is not proof they ran the same thing. Check the version suffix first, then treat a couple of points of unexplained spread as normal.
The October cards appear to settle one of those disputes and open several more. DeepSeek’s V4.1-Flash card now prints 43.2 on AutomationBench for V4 Pro, the Z.ai figure, where its own V4 Pro card printed 31.8. The 1017 card labels the suite v1.0.6, so DeepSeek has moved to the versioned split. The base V4.1-Flash card gives Kimi K3 59.8 on HLE with tools, agreeing with Z.ai, though it lists GLM-5.3’s AutomationBench at 48.8 against Z.ai’s own 48.2. Meta’s Muse card goes the other way: it gives K3 56.0 on that row, and it puts GLM-5.3 at 83.9 on Terminal-Bench 2.1 against 88.2 on Z.ai’s own card, with the harness named as Terminus 2. GLM-5.3’s GPQA Diamond is 91.7 on Meta’s card and 88.1 on DeepSeek’s, a 3.6 point gap with no version label on either side. GDPval-AA v2 drifts furthest: Meta’s card gives GLM-5.3 1653 and Kimi K3 1537, against 1769 and 1682 on Z.ai’s. Treat any cross-vendor table, including the one above, as the publishing vendor’s run under its chosen harness.
License Comparison
Licensing is where “open source” gets complicated. Some models are truly permissive (Apache 2.0, MIT), while others come with usage caps, geographic restrictions, or revenue gates. Read the fine print before building a product on any of these. The licensing trend reversed in August. Google had already moved Gemma 4 to plain Apache 2.0, dropping the separate agreement Gemma 3 required, and Moonshot had shipped Kimi K3 under a custom license most people assumed was MIT. Then Alibaba and Z.ai both split their own lineups, keeping the permissive license on the smaller models and putting custom terms on the flagship. October added two more custom licenses: Meta’s Muse Model License and Liquid AI’s LFM Open License, while DeepSeek and OpenBMB stayed on MIT and Apache 2.0.
| License | Models | Commercial Use | Key Restrictions |
|---|---|---|---|
| Apache 2.0 | Qwen3.8-27B, Qwen 3/3.5/3.6 (all open releases), gpt-oss-120b/20b, Gemma 4 (all sizes), MiniCPM-V 4.7, Muse Glimmer 30B, Mistral Large 3, Mistral Small 4, Mixtral, Mistral 7B | Yes, unrestricted | None (gpt-oss and Muse Glimmer add a usage policy) |
| MIT | DeepSeek R1/V3.2/V4 Pro/V4 Flash/V4.1-Flash (including the 0813, 1017 and Flash Vision releases), GLM-5, GLM-5.2, GLM-5.3-Flash, Phi-4 (all variants) | Yes, unrestricted | None |
| Modified MIT | Kimi K2 | Yes | Attribution required for very large commercial deployments |
| Kimi K3 License | Kimi K3 | Yes, with gates | Custom license, not MIT: model-as-a-service providers above $20M/year revenue need a separate agreement, and products above 100M MAU or $20M/month revenue must display “Kimi K3”. Internal use exempt |
| Qwen3.8-Max License | Qwen3.8-Max (2.4T-A95B) | Yes, with gates | If the licensee and its affiliates take more than $50M in aggregate revenue over any consecutive 12 months and run a model-as-a-service or AI work assistant business, a separate license from Qwen is required before any commercial use, not only for the resale itself. Products above 100M monthly active users or $20M monthly revenue must display the model name. Internal use is exempt only where the model, its outputs, and its capabilities stay away from third parties |
| Qwen Community License 1.0 | Qwen3.8-Flash-Next | Yes, with gates | Any model-as-a-service or AI work assistant business needs a separate license, with no revenue threshold at all. Same display requirement above 100M monthly active users or $20M monthly revenue. Internal use is exempt on the same third-party condition |
| GLM-5.3 License | GLM-5.3 | Yes, with gates | Model-as-a-service operators whose revenue passes $10B in any 12 months must pass a Z.ai security review before commercial use. No display requirement |
| Muse Model License | Muse Spark Open | Yes, with gates | Every product or service built on the model or any derivative must display “Built with Muse Spark Open”, at any size. Derivatives include models distilled from its outputs. Model-as-a-service businesses need a separate Meta agreement once the licensee and its affiliates pass $20M gross revenue in 12 months; internal use, research and features inside your own non-MaaS product are exempt from that clause. The usage policy says it is not intended for under-18s, and deployers whose systems minors can reach must assess and mitigate that risk |
| LFM Open License v1.0 | LFM2.5-24B-A2B | Only under $10M annual revenue | Apache 2.0-style redistribution terms, but commercial use is licensed only while your legal entity stays under $10M annual revenue. On the license’s definition of commercial use (any direct or indirect commercial advantage), internal use at a for-profit company counts, and the revenue test covers your whole group because affiliates under common control are part of the legal entity. Non-commercial and research use is not gated |
| Llama 4 Community | Llama 4 Scout, Llama 4 Maverick | Yes, free under 700M MAU | EU multimodal restrictions; Meta license required above 700M monthly active users |
| Llama 3.3 Community | Llama 3.3 70B | Yes, free under 700M MAU | Same MAU threshold as Llama 4 |
| Gemma | Gemma 3 (all sizes) | Yes (requires agreement) | Must accept Google’s terms; Gemma 4 dropped this requirement entirely |
| CC-BY-NC | Command A, Command R+ | No | Non-commercial only; separate agreement required for commercial deployment |
| TII Falcon-LLM 2.0 | Falcon 3 (all sizes) | Yes | Apache 2.0-based with an acceptable use policy; no royalty or revenue gate |
Which license gate applies to you
For five of the seven licenses below, the gates only bite in two situations: reselling inference, or running a consumer product at real scale. Everything else, including internal deployment and fine-tuning, stays free under them. The two October licenses break that pattern. LFM gates any commercial use by company revenue, and Muse asks for attribution from the first user. Apache 2.0 and MIT attach no conditions at all, while the two Qwen licenses exempt internal use only where the model, its outputs, and its capabilities stay away from third parties. The thresholds come from the LICENSE files in each repository, including the Qwen3.8-Max license text.
| License | Model | Trigger for a separate agreement | Attribution required |
|---|---|---|---|
| Apache 2.0 | Qwen3.8-27B | None | No |
| MIT | GLM-5.3-Flash, DeepSeek V4 | None | No |
| Qwen3.8-Max | Qwen3.8-Max | Any commercial use once licensee plus affiliates pass $50M / 12 months and run inference resale or an AI work assistant business | Above 100M MAU or $20M/month |
| Qwen Community 1.0 | Qwen3.8-Flash-Next | Any inference resale or AI work assistant business, no threshold | Above 100M MAU or $20M/month |
| GLM-5.3 | GLM-5.3 | Inference resale above $10B / 12 months | No |
| Muse Model License | Muse Spark Open | Inference resale once licensee plus affiliates pass $20M / 12 months | Always: “Built with Muse Spark Open” |
| LFM Open License v1.0 | LFM2.5-24B-A2B | Any commercial use once your entity reaches $10M annual revenue | No display line; keep the license and notices |
The practical asymmetry is between the two Qwen licenses. Qwen3.8-Max lets a hosting business run until the licensee and its affiliates clear $50M in aggregate revenue, while the Qwen Community License covering Qwen3.8-Flash-Next requires a separate agreement for any inference resale at all, with no revenue floor. The smaller, cheaper model carries the stricter redistribution terms. Z.ai’s $10B threshold is high enough that it reaches only hyperscalers, so GLM-5.3 is effectively permissive for everyone reading this, and GLM-5.3-Flash stays plain MIT.
Muse reuses Kimi K3’s $20M resale figure but drops the scale floor on attribution. K3 asks for its name only above 100M monthly users or $20M a month in revenue, while Muse asks every product and every model distilled from its outputs. LFM is the license to read twice if you work for a mid-size company. A 24B model built for on-device agents is exactly what an internal team would deploy, and above $10M in revenue that deployment falls outside the grant.
For most self-hosting, the field is still wide open: Qwen3.8-27B and Gemma 4 under Apache 2.0, DeepSeek V4 and GLM-5.3-Flash under MIT, all fine-tunable and deployable commercially with zero royalties. Even OpenAI’s gpt-oss models ship under Apache 2.0, with a usage policy attached. MiniCPM-V 4.7 joins that group under Apache 2.0, and DeepSeek V4.1-Flash keeps MIT. Read the Kimi K3, Muse and Qwen Community terms carefully if you are building a product that resells inference, and note that the Llama licenses still carry the 700M MAU cap plus the EU multimodal exclusion.
Self-Hosting Resource Requirements
Benchmarks don’t tell you how a model feels when you’re actually running it. We tested six popular small models on a modest Ubuntu 24.04 VM (4 vCPUs, 16 GB RAM, CPU-only inference) using Ollama in March 2026. Each model answered the same prompt to keep things consistent.
| Model | Ollama Tag | Disk Size | RAM Usage | Response Time (CPU) | Notes |
|---|---|---|---|---|---|
| Llama 3.2 3B | llama3.2:3b | 2.0 GB | 11.4 GB | 88s | Clear, well-structured responses |
| Gemma 3 4B | gemma3:4b | 3.3 GB | 4.2 GB | 94s | Clean, structured, concise output |
| Phi-4 Mini 3.8B | phi4-mini | 2.5 GB | 8.9 GB | 97s | Good reasoning, occasional formatting artifacts |
| Mistral 7B | mistral:7b | 4.4 GB | 7.4 GB | 125s | Concise, accurate |
| Qwen 3 8B | qwen3:8b | 5.2 GB | 5.8 GB | 433s | Thinking mode adds latency; strong final answers |
| DeepSeek R1 8B | deepseek-r1:8b | 5.2 GB | 5.8 GB | 433s | Chain-of-thought reasoning; slow on CPU |
Gemma 3 4B stands out for RAM efficiency at just 4.2 GB, making it the best fit for memory-constrained environments. The reasoning models (Qwen 3 8B and DeepSeek R1 8B) took over 7 minutes each because their chain-of-thought process generates far more tokens before producing a final answer. On a GPU, those times drop to seconds. If you are choosing hardware for local inference, our GPU buyer guide for local LLMs breaks down which card fits each model size, and the VRAM requirements guide maps parameter counts to memory directly.
What the 2026 flagships actually need
The frontier open models are a different class of problem, so in August 2026 we measured two of them ourselves instead of quoting vendor claims. The gpt-oss and Muse Spark Open rows are vendor figures we have not run.
| Model | Smallest usable quant | Memory floor | Measured speed | Hardware we tested on |
|---|---|---|---|---|
| DeepSeek V4 Flash | UD-IQ3_XXS, 103 GB | ~112 GB RAM (measured peak) | 5.61 tok/s generation, 12.33 tok/s prompt processing | 8-core CPU server, 128 GB RAM |
| Kimi K3 | UD-IQ1_S, 594 GB | 610 GB RAM + VRAM combined | 0.10 tok/s generation | 4x A100 40GB + 2 TB RAM |
| gpt-oss-120b | MXFP4 (ships natively quantized) | Single 80 GB GPU (vendor figure) | Not benchmarked by us | H100 or MI300X class |
| Muse Spark Open | NVFP4, 533 GB download | 700 GB VRAM (vendor figure; 2 TB for BF16) | Not benchmarked by us | GPU node with 700 GB+, or two Mac Studio M5 Ultra with MLX (vendor) |
Two findings from those runs are worth internalizing. First, DeepSeek V4 Flash is genuinely runnable on a 128 GB machine, and prompt processing, not generation, is the wall: at 12.33 tok/s prefill, a 10,000-token prompt takes over 13 minutes before the first output token appears. The full methodology and every measurement is in our DeepSeek V4 Flash local hardware test. Second, Kimi K3 is not practically self-hostable: one token every 10 seconds on four A100s, because 594 GB of weights turn the run into a memory-bandwidth problem no consumer budget solves. The numbers are in our Kimi K3 local run write-up. If a 100 GB class model is your target, the realistic home options are a 128 GB unified-memory box (see our Ryzen AI Max+ 395 mini PC comparison) or a multi-GPU workstation; for everything smaller, a mini PC built for local AI covers the 4B to 30B range comfortably.
Pull and run any of the small models with a single command:
ollama run gemma3:4bCheck our Ollama commands cheat sheet for the full list of management commands, and the Ollama models cheat sheet for current tags and quantization options across every family on this page.
What the weights actually weigh
Parameter count stopped predicting download size in 2026. Vendors now publish at four different precision mixes, so the model with the most parameters on this page is not the biggest download. Every figure below is the sum of the safetensors shards in each repository, taken from the repository’s own index rather than estimated from the parameter count.
| Model | Parameters | Repo precision | Download | Bytes per param |
|---|---|---|---|---|
| Qwen3.8-Max (2.4T-A95B) | 2.45T | BF16 | 4,892 GB | 2.00 |
| Muse Spark Open | 822B stored (816B headline) | BF16 | 1,644 GB | 2.00 |
| Kimi K3 | 2.78T | Pre-quantized | 1,561 GB | 0.56 |
| GLM-5.2 | 753B | BF16 | 1,507 GB | 2.00 |
| DeepSeek V4 Pro (0813) | 1.65T | Pre-quantized | 893 GB | 0.54 |
| GLM-5.3 | 753B | FP8 | 756 GB | 1.00 |
| Muse Spark Open NVFP4 | 816B (484B stored, packed) | NVFP4 | 533 GB | 0.65 |
| DeepSeek V4.1-Flash (1017) | 763B stored | FP8 + FP4 experts | 510 GB | 0.67 |
| Qwen3.8-Flash-Next | 180B | BF16 | 360 GB | 2.00 |
| GLM-5.3-Flash | 320B | FP8 | 328 GB | 1.02 |
| DeepSeek V4 Flash (0731) | 304B stored (284B headline) | Pre-quantized | 167 GB | 0.55 |
| gpt-oss-120b | 117B | MXFP4 | 65 GB | 0.56 |
| MiniCPM-V 4.7 35B-A3B | 35.2B | BF16 | 70 GB | 2.00 |
| Muse Glimmer 30B | 29.8B stored | BF16 | 60 GB | 2.00 |
| Qwen3.8-27B | 27.8B | BF16 | 56 GB | 2.00 |
| LFM2.5-24B-A2B | 24.0B | BF16 | 48 GB | 2.00 |
Read that table by the last column and the pattern is clean: BF16 repos land at 2.00 bytes per parameter, FP8 repos at 1.00, and the pre-quantized 4-bit-class repos at roughly 0.55. The consequences are counterintuitive. Kimi K3 carries more parameters than Qwen3.8-Max (2.78T against 2.45T) and downloads at less than a third the size. DeepSeek V4 Pro holds more than twice the parameters of GLM-5.2 and still downloads smaller, 893 GB against 1,507 GB, because DeepSeek ships quantized weights while the default GLM-5.2 repository is BF16. Z.ai does publish a GLM-5.2 FP8 build, and it is byte-identical in size to GLM-5.3’s default FP8 repository, which makes the point sharper: what changed between those two releases is which precision Z.ai put in front of you, not how big the model is.
The October releases add a fourth band. DeepSeek V4.1-Flash stores its non-expert weights in FP8 and its experts in FP4, which lands at 0.67 bytes per stored parameter, and Meta’s NVFP4 build of Muse Spark Open sits at 0.65 against its 816B headline. In BF16, Muse at 816B downloads bigger than Kimi K3 at 2.78T, 1,644 GB against 1,561 GB, which is the Qwen3.8-Max lesson again: the published precision decides the download, not the parameter count.
GLM-5.3 is the clearest illustration. Z.ai’s model card states that it uses the same base model as GLM-5.2 and that every gain comes from post-training, and the config files bear that out: identical layer count, hidden size, expert count, attention heads, vocabulary, and context length. The only structural difference is that the default GLM-5.3 repository ships FP8, which is why the same 753B parameters download as 756 GB instead of 1,507 GB. Z.ai publishes a BF16 build as a separate repository, and it comes out at 1,507 GB, matching GLM-5.2 exactly. Same model, half the bytes, different license.
One trap to avoid if you script this yourself: multiplying a parameter count by two bytes is no longer a safe estimate, and neither is multiplying by one. The gpt-oss-120b tensors are declared as unsigned 8-bit, which implies about 119 GB once its BF16 tensors are added in, but the shard index totals 65 GB because the 4-bit weights are packed two to a byte. Read the shard index, or you will provision storage that is off by a factor of two in either direction.
Qwen 3.8, 3.6 and Earlier (Alibaba)
Qwen is arguably the most versatile open-source model family available, and the release cadence is relentless: four generations in sixteen months. The blanket Apache 2.0 guarantee ended with Qwen3.8, which now splits licensing by model size. The Qwen 3 series (April 2025) spans from a 0.6B edge model to a 235B MoE flagship. Qwen 3.5 (February 2026) made the family natively multimodal, pushed context to 256K, and extended language coverage to 201 languages.
Qwen 3.6 (April 2026) is the generation that changed the economics. The two open releases pair a hybrid attention design (Gated DeltaNet linear attention mixed with standard self-attention) with a native 262K context extensible to roughly 1M tokens via YaRN scaling, and both accept text, images, and video. Qwen3.6-35B-A3B activates just 3B parameters per token yet posts 85.2% MMLU-Pro and 73.4% SWE-bench Verified. The dense Qwen3.6-27B is the headline act: 77.2% on SWE-bench Verified beats Alibaba’s own 397B MoE flagship, and the Q4_K_M quant is 16.8 GB, which fits a single 24 GB consumer GPU but overshoots a 16 GB card before any context is allocated. Both introduce Thinking Preservation, which carries reasoning traces across conversation turns instead of discarding them, a meaningful win for agent workloads.
The toggleable thinking mode remains the family’s signature feature on every size except the flagship. Enable chain-of-thought for math, logic, and coding; disable it for fast responses on straightforward queries. Qwen3.8-27B and Flash-Next still accept enable_thinking: False, so one model serves both use cases. Qwen3.8-Max does not: its card states thinking cannot be disabled at all.
Qwen3.8 (August 2026) is where the family stops being uniform. Three open releases now sit under three different licenses. Qwen3.8-27B is a dense 27.8B native vision-language model, still Apache 2.0, and the most practical of the three at 56 GB in BF16 or 31 GB in the FP8 build. It reports 84.3 on OSWorld-Verified for computer use against 72.7 for the Opus 4.6 Max column on the same card, and 90.3 on LiveCodeBench v6. Qwen3.8-Flash-Next is a 6B-active MoE under the Qwen Community License, and its own card lists an unusual shape: 125B parameters with 6B activated, plus a separate 51B n-gram embedding table and a 4B MTP head, which is how the repository reaches 180B on disk rather than the headline figure.
Qwen3.8-Max, published as Qwen3.8-2.4T-A95B, is the 2.45 trillion parameter flagship with 95B active across 512 experts. Two things about it catch people out. The open weights are text-only and thinking cannot be disabled, per the model card, so every response carries a reasoning block whether you want one or not. The hosted “Qwen3.8-Max” on Qwen Cloud is a different product that adds vision input, a non-thinking mode, and 1M context by default. And at BF16 the weights are 4,892 GB, so this is a model you rent rather than run.
gpt-oss (OpenAI)
OpenAI shipping open weights was the event nobody had on their 2025 card. gpt-oss-120b and gpt-oss-20b, released August 2025 under Apache 2.0, are the company’s first open-weight models since GPT-2. Both are MoE reasoning models with a 128K context and three configurable reasoning effort levels (low, medium, high), and both ship natively quantized in MXFP4.
The efficiency story is the point. gpt-oss-120b activates only 5.1B of its 117B parameters per token and runs on a single 80 GB GPU, delivering near-parity with OpenAI’s o4-mini on core reasoning benchmarks and 62.4% on SWE-bench Verified at high reasoning effort. gpt-oss-20b activates 3.6B of 21B and runs on machines with 16 GB of memory, which puts a capable reasoning model on ordinary laptops. Both are text-only. A year on, community tooling support is excellent across Ollama, llama.cpp, and vLLM.
Kimi K2 and K3 (Moonshot AI)
Moonshot AI escalated the open-weights race twice. Kimi K2 (July 2025) is a 1T parameter MoE with 32B active and a Modified MIT license whose only catch is an attribution requirement for very large deployments; the original shipped with a 128K context, extended to 256K in later refreshes. It earned its reputation on agentic and tool-use workloads.
Kimi K3, whose weights went public in late July 2026, is the largest open model ever released: 2.8T total parameters, 104B active from 896 experts, native vision, video input, and a 1M token context. The weights arrive as 96 shards totaling about 1.56 TB on Hugging Face. Benchmark-wise it held the top open-weight spot on the aggregate indexes until GLM-5.3 passed it. Two caveats matter before you get excited. The license is a custom Kimi K3 document, not MIT: model-as-a-service providers above $20M/year revenue need a separate agreement, and very large products must display “Kimi K3” attribution, though internal use stays unrestricted. And self-hosting is effectively off the table: our own run needed 610 GB of combined RAM and VRAM for the smallest quant and generated at 0.10 tokens per second on four A100s. Full details in the Kimi K3 local run test. One operational trap from using the hosted version: K3 defaults to very high reasoning effort, and a low max-token budget can be consumed entirely by hidden reasoning, returning an empty response you still pay for. Budget 5,000+ output tokens or lower the reasoning effort.
DeepSeek V4, V3.2, and R1
DeepSeek V4 arrived in April 2026 as two MIT-licensed models sharing a sparse-attention architecture (token-wise compression plus DeepSeek Sparse Attention) and a 1M token default context: V4 Pro at 1.6T total parameters with 49B active, and V4 Flash at 284B total with 13B active. Both support thinking and non-thinking modes. Both lines have since been re-post-trained. For Flash weights, target the DeepSeek-V4-Flash-0731 repository, which superseded the original preview. V4 Pro got the same treatment on 13 August 2026 as DeepSeek-V4-Pro-0813, but that release is more than post-training. The transformer backbone is unchanged, and DeepSeek’s card describes it as built on the V4 Pro preview structure with a DSpark speculative decoding module attached. The config adds four dspark_* keys the original does not carry, and the parameter count grows from 1.599T to 1.650T, so the extra 51.7B parameters are the draft model living inside the same checkpoint. That is why you enable it with a speculative-algorithm flag and no separate draft-model path. An experimental multimodal build, DeepSeek-V4-Flash-Vision-Exp, followed on 31 August: it adds a vision encoder and aligner to the Flash architecture, reports gains on multimodal agent tasks such as ApexBench (36.5 against 26.2), and holds text agent performance roughly level. Read that ApexBench delta with the card’s own footnote in mind: the 26.2 baseline is text-only Flash ignoring the images in the input, so it measures having vision at all rather than one vision model beating another.
DeepSeek V4.1-Flash is not a refresh of V4 Flash, whatever the name suggests. Its card describes a new base trained from scratch on 45T multimodal tokens, with 552B backbone parameters against V4 Flash’s 284B. A causal encoder-decoder layout activates 8B parameters per token during prefill and 16B during decode, images are a native input, and 196B parameters of Engram conditional memory are looked up by token. The headline is the KV cache, which DeepSeek puts at 890 bytes per token, about a quarter of V4 Flash’s. The base repository dates from 10 September 2026. DeepSeek-V4.1-Flash-1017 followed on 17 October as a minor update with the same config and the same 510 GB on disk, and it reports 42.6 on HLE against 36.8 and 77.0 on DeepSWE v1.1 against 74.2.
For self-hosting the trade is size: 510 GB of FP8 and FP4 weights against 167 GB for V4 Flash 0731, so the 128 GB box that runs V4 Flash cannot hold V4.1-Flash at the precision DeepSeek publishes. Two more changes bite if you serve it yourself. Reasoning effort is now a budget from 1 to 100. The card says no Jinja chat template is released and points to a Python reference encoder and the deepseek-recipe libraries, but the base repository has carried a merged chat_template.jinja since 1 October that accepts either the integer or low, medium, high and max (mapped to 50, 62, 75 and 100) and defaults to high. The 1017 repository has no template file, and its config is identical, so take the template from the base repository when you serve the update.
V4 Pro’s 80.6% on SWE-bench Verified at max thinking is the highest open-weight score published on that specific benchmark, and the qualifier matters more than it used to: not one of the model cards released in August 2026 reports SWE-bench Verified at all. Z.ai and DeepSeek dropped SWE-bench altogether in favour of Terminal Bench, DeepSWE, and NL2Repo, and Alibaba kept only the Pro, Multilingual, and Multimodal splits, so coding scores from this generation cannot be lined up against the previous one. V4 Flash is the one you can realistically run yourself: we measured 5.61 tokens per second generation on a CPU-only 128 GB box with the 103 GB UD-IQ3_XXS quant. Setup, expert-offload flags, and the prompt-processing bottleneck are covered in our hands-on DeepSeek V4 Flash guide.
The older generation still earns its keep. R1’s 97.3% on MATH-500 remains near the top, and the distilled versions (1.5B through 70B) put genuine chain-of-thought reasoning on consumer hardware; see running DeepSeek R1 locally with Ollama. V3.2 (December 2025) pioneered thinking integrated into tool-use workflows, and its Speciale variant achieved gold-medal performance at IMO 2025, IOI 2025, and the ICPC World Finals.
GLM-5.3, GLM-5.2 and GLM-5 (Z.ai)
Z.ai (formerly Zhipu AI) released GLM-5 in February 2026: 744B total parameters, 40B active, MIT license, trained entirely on 100,000 Huawei Ascend 910B chips with no US-manufactured hardware. Its 77.8% on SWE-bench Verified stood as the open-weight coding record for months.
GLM-5.2, launched June 13, 2026 with weights on Hugging Face days later, pushes to 753B total with roughly 40B active, a 1M token context, and what was the strongest open coding profile on this page until GLM-5.3 arrived: 62.1 on SWE-bench Pro, 81.0 on Terminal-Bench 2.1, 99.2 on AIME 2026, and 91.2 on GPQA Diamond, all from GLM-5.2’s own card. GLM-5.3 republishes only the Terminal-Bench 2.1 line of those four, at 88.2 against GLM-5.2’s 81.0, so the like-for-like comparison runs through the sixteen rows of its own card instead, where it beats GLM-5.2 in every one. Weights ship under MIT. Know what you are getting, though: GLM-5.2 is text-only. For most of the year Z.ai’s vision line (GLM-5V-Turbo) was API-only with no downloadable weights. GLM-5.3-Flash changed that in August.
GLM-5.3 and GLM-5.3-Flash both landed in late August 2026, and they are different kinds of release. GLM-5.3 is not a new base model. Z.ai’s GLM-5.3 model card states plainly that it uses the same base as GLM-5.2 and that every gain comes from post-training, which the config files confirm down to the layer count, expert count, attention heads, and vocabulary. What changed is the packaging and the terms: the default repository ships FP8 rather than BF16, halving the download to 756 GB, and the license moved off MIT to Z.ai’s own GLM-5.3 terms with a $10B security-review clause for inference resellers.
GLM-5.3-Flash is the more interesting one for self-hosting. It is the first natively multimodal model in the GLM-5 series, 320B total with 18B active, and it keeps the MIT license the flagship gave up. Z.ai built it on a new base with a hybrid sparse and linear attention design and says it outperforms GLM-5.2 across benchmarks and real-world workloads at roughly a tenth of the price. At 328 GB in FP8 it is also the first GLM flagship-class release that fits inside the 640 GB of an eight-way 80 GB GPU node, which the 756 GB GLM-5.3 does not. Size the node on more than the weights, though: at the 1M context it advertises, the KV cache needs room alongside those 328 GB.
Gemma 4 and Gemma 3 (Google)
Gemma 4 launched March 31, 2026 with four sizes, and fixed the two biggest complaints about the family in one release. The license is now plain Apache 2.0, no separate Google agreement required. And the lineup finally includes an MoE option: the 26B A4B activates 3.8B parameters from 128 experts, giving near-flagship quality at small-model inference cost. The encoder-free 12B followed in June 2026, rounding out the family to five sizes.
The sizes split cleanly by deployment target. E2B and E4B use per-layer embeddings to bring effective parameter counts (2.3B and 4.5B) below their raw sizes, run a 128K context, and take text, image, and audio input; they are the phone and edge-device models. The 12B has the interesting architecture: encoder-free, projecting image and audio directly into the transformer’s embedding space instead of bolting on separate encoders. The 26B A4B and 31B dense flagships run a 256K context with text and image input. All sizes support function calling, structured JSON output, and native system instructions, and training covered over 140 languages.
Gemma 3 remains relevant for one reason: it is everywhere, and it is proven. Our March CPU tests above show why the 4B stays popular in tight-memory deployments. New projects should start on Gemma 4.
Muse Spark Open and Llama 4 (Meta)
Muse Spark Open is Meta’s first large open-weight release since Llama 4. It shares the meta-models Hugging Face organization with Muse Glimmer 30B, the dense model Meta published in August, but it ships under a license of its own. That organization carries Hugging Face’s verified badge under the name Meta Inc. The card credits Meta Superintelligence Labs and dates the release 15 October 2026. It is an 816B mixture-of-experts with 42B active, 256 experts routed 8 at a time across 62 layers, a 1M token context, and a 1.8B perception encoder, with text, image, video and audio input. Access is gated: you fill in a form with your full legal name, date of birth and organization, and Meta approves it by hand.
Meta’s own benchmark table is honest about where it lands. Muse leads the open field on MCP-Atlas (89.2), JobBench (61.8) and HLE-Diamond without tools (26.1). It trails Kimi K3 on DeepSWE (60 against 68) and Terminal-Bench 2.1 (76.4 against 85.0), and its Artificial Analysis Intelligence Index of 39 sits below GLM-5.3, Kimi K3 and Qwen3.8-Max. The outlier is safety. On the StrongREJECT adaptive jailbreak test, where lower is better, Muse scores 0.5 and every other open model on the card scores between 62.9 and 86.4, which is the extra safety post-training the card describes. If you want an open model that refuses hard by default, that column is the reason to look at Muse. If you want a coding agent, the numbers point at GLM-5.3 or Kimi K3. Hardware is the other filter: the card asks for 700 GB of GPU memory for the NVFP4 build and 2 TB for BF16, or two Mac Studio M5 Ultra machines running MLX.
Muse Glimmer 30B is the Meta model most readers can actually run. It is a dense 29.6B model, perception encoder included, distilled from Muse Spark, with text and image input, a 131,072+ token context and plain Apache 2.0 terms, so none of the Muse Model License conditions apply to it. The BF16 repository is 60 GB, and Meta’s card targets a 24 GB card with a 17 GB K-quant that it measures at 1.0% average degradation; add the 1.4 GB vision projector and the 1.6 GB DFlash drafter and you are near 20 GB before the KV cache. The card also reports 74.9 tokens per second on an RTX 5090 under llama.cpp, rising to 233.4 with the bundled DFlash drafter, and puts it at 76.0 on SWE-bench Verified against 77.2 for Qwen3.6-27B on the same table. Reasoning strength is set in the system prompt (Reasoning strength: high) rather than through a request parameter.
Meta’s Llama 4 family (April 2025) brought MoE to the Llama line. Scout (109B total, 17B active) carries a 10 million token context window, still the longest of any open model by a wide margin. Maverick (400B total, 17B active, 128 experts) targets production deployments with a 1M context. Both are natively multimodal.
The Llama line itself has been static since. Behemoth, the roughly 2T parameter flagship Meta paused in 2025 over incremental internal results, still has no weights in the meta-llama organization in October 2026. No “Llama 4.5” or “Llama 5” has appeared in Meta’s Hugging Face organizations either. Meta’s newer open models carry the Muse name instead: Glimmer 30B in August and Spark Open in October. Meanwhile the licensing catch is unchanged: the Llama 4 Community License is free under 700M MAU, but the Acceptable Use Policy excludes multimodal rights for EU-based individuals and companies, which effectively restricts the whole Llama 4 family in the EU. Llama 3.3 70B (dense, text-only) is unaffected and remains popular because it needs no MoE-aware serving infrastructure. One GPU means a big one, though: a 70B model wants 40 GB or more at 4-bit, so a 32 GB RTX 5090 cannot load it while a 96 GB card can. Our RTX PRO 6000 vs RTX 5090 local AI benchmarks measured Llama 3.1 70B at 34.2 tokens per second on the PRO 6000 and found the 5090 unable to load it at all.
Mistral Models
Mistral’s current pair covers both ends of the spectrum under Apache 2.0. Mistral Large 3 (December 2025) is a 675B MoE with 41B active, multimodal across 80+ languages. Mistral Small 4 (March 2026) is the clever one: 119B total with only 6B active through 4-of-128 expert routing, unifying instruction following, configurable-depth reasoning, and multimodal input in one model. It absorbed what used to be three separate product lines (Magistral for reasoning, Pixtral for vision, Devstral for agentic coding).
A new open-weight family is coming: Mistral confirmed in early July 2026 that a larger “fat but sparse” MoE family entered early access with research and government partners, with no public parameter count or ship date. Benchmarks have since surfaced without weights: Meta’s Muse card scores a “Mistral Large 4” against its own model, and Artificial Analysis lists a Mistral Large 4 Preview on 38, but no Large 4 weights sit in Mistral’s Hugging Face organization as of October 2026. The original Mistral 7B and Mixtral 8x7B remain in wide deployment and are still serviceable Apache 2.0 fallbacks on old hardware, though every current alternative outperforms them.
Phi-4 (Microsoft)
Microsoft’s Phi-4 family proves that smaller models can punch above their weight on specific tasks. The 14B base model and its reasoning variants excel at math and logic, consistently outperforming larger models there. Phi-4 Mini at 3.8B with a 128K context remains one of the best options for resource-constrained deployments that need long context, though in our testing it used 8.9 GB of RAM, more than expected for its size, likely due to context window allocation.
Phi-4 Reasoning Vision (15B, March 2026) adds image understanding to the reasoning pipeline via a SigLIP-2 vision encoder, with dynamic resolution up to 3,600 visual tokens for GUI grounding and document analysis. It decides per query whether deep reasoning is warranted, saving compute on simple questions. All Phi-4 variants ship under MIT.
LFM2.5-24B-A2B and MiniCPM-V 4.7
These are the two October releases built for a single consumer GPU or a laptop, and both use a small active slice of a mid-size MoE. Neither has been through our lab yet, so the sizes below come from their repositories and the speed figures from their cards.
Liquid AI’s LFM2.5-24B-A2B has 24B parameters with 2.45B active and a 131,072 token context. Of its 40 layers, 30 are gated short-convolution blocks and 10 are grouped-query attention, and each MoE block routes 4 of 64 experts. It is text-only and always reasons: the chat template opens every answer with a <think> tag, which its 24B predecessor did not do. Liquid pitches it for on-device agents and tool calling, quotes 184 tokens per second on an Apple M5 Max, and recommends pairing it with retrieval for knowledge-heavy work instead of trusting its own recall. The Hugging Face repository was created on 12 October 2026, and the card gives no separate release date.
The size is what makes it interesting. Liquid’s Q4_K_M GGUF is 14.5 GB, and Ollama lists the same build as lfm2.5:24b at 15 GB, so a 16 GB card holds the weights with about 2.5 GB to spare for the KV cache and runtime overhead. That is a usable short-context agent on an RTX 5070 Ti or a 5060 Ti 16GB, but not a 131K one with an f16 KV cache. Our RTX 5070 Ti vs 5060 Ti 16GB benchmarks show what those two cards do with models from 4.9 to 15 GB, and the bandwidth gap between them, 896 against 448 GB/s, is what sets generation speed once a model fits.
OpenBMB’s MiniCPM-V 4.7 35B-A3B is the first MoE in the MiniCPM-V line: 35.2B parameters, 3B active, single-image, multi-image and video input, Apache 2.0. Its config.json declares a qwen3_5_moe_text backbone with 256 experts routed 8 at a time and 262,144 positions; the card itself states no context length. The card’s headline is 13+ tokens per second on an iPhone 17 Pro Max with 3 GB of memory. That figure comes from Lookahead Expert Scheduling in OpenBMB’s own llama.cpp fork, which streams experts from storage instead of holding them all, so read it as a phone demo rather than a memory floor. A 1B variant ships alongside it for smaller devices.
On a GPU the published file sizes decide it, and they map onto card sizes like this:
- 16 GB: LFM2.5 Q4_K_M (14.5 GB), or MiniCPM-V 4.7 IQ3_XS (14.5 GB) plus its 0.9 GB 4x vision projector
- 24 GB: MiniCPM-V 4.7 Q4_K_M (21.2 GB) plus the 1.1 GB vision projector
- 64 GB and up: LFM2.5 in BF16 (48 GB)
- 96 GB: MiniCPM-V 4.7 in BF16 (70 GB)
If your card has 16 GB, LFM2.5 is the one that fits at 4-bit, and it is the stronger pick for a tool-calling agent as long as your company stays under the $10M license threshold. If you need vision on the same card, MiniCPM-V 4.7 only fits as the IQ3_XS build with the smaller vision projector, because llama.cpp loads that projector file on top of the language weights. Either accept the 3-bit loss or move to a 24 GB GPU and run its Q4_K_M build.
Command A, Falcon 3, and the Rest
Command A (Cohere, 111B dense) offers a 256K context, runs on two A100/H100 GPUs, and stays optimized for retrieval-augmented generation and tool use. The CC-BY-NC license limits it to non-commercial use without a separate agreement, which keeps it out of most production shortlists despite the solid engineering. Cohere’s smaller releases, Tiny Aya (3.35B, 70+ languages, CC-BY-NC) and the Apache 2.0 Cohere Transcribe speech model, fill niches the big families ignore.
Falcon 3 from TII Abu Dhabi (1B to 10B, trained on 14T tokens) ships under the TII Falcon-LLM License 2.0, an Apache 2.0-based license with an acceptable use policy and free commercial use. DBRX and Grok-1, both early-2024 MoE releases, have aged out of practical relevance, and while xAI published Grok 2 weights in August 2025, the promised Grok 3 weights are still absent from its Hugging Face organization in October 2026.
Thinking budgets, and why the default costs you money
Every flagship released since August ships a thinking-budget control, and on Qwen3.8, GLM-5.3 and Muse Spark Open the default is the most expensive setting. Serve one of these without passing the parameter and you pay for maximum deliberation on every request, including the trivial ones. Four vendors now use the same parameter name, and their named levels overlap only on low, while DeepSeek V4.1-Flash takes a 1 to 100 budget underneath its names.
| Model family | Parameter | Accepted levels | Default |
|---|---|---|---|
| Qwen3.8 (all sizes) | reasoning_effort | low, medium, xhigh | xhigh |
| GLM-5.3, GLM-5.3-Flash | reasoning_effort | low, high, max | max |
| DeepSeek V4 Pro | reasoning_effort | low, high, max | Not documented on the model card |
| DeepSeek V4.1-Flash | reasoning_effort | Integer 1 to 100, or low, medium, high, max through the base repo’s template | high (75) in that template; card benchmarks run at 100 |
| Muse Spark Open | reasoning_effort | low, medium, high, xhigh | xhigh |
| Gemini 3.8 Flash (hosted, for contrast) | thinkingLevel | low, medium, high | medium |
Three details in that table cause real bugs. Z.ai documents that GLM-5.3 falls back to max when the value is missing or unrecognized, so a typo does not raise an error, it silently buys you the most expensive setting. Meta went the opposite way: the Muse card says any value outside its four levels returns an error, so the same wrapper that silently overspends on GLM fails loudly on Muse. And max is not a level Qwen3.8 or Muse defines at all, while DeepSeek V4.1-Flash maps its names onto a numeric budget, so a wrapper that hardcodes one vendor’s vocabulary is not portable to any of the others.
Qwen3.8-Max goes further and removes the choice: the model card states that thinking cannot be disabled and every response begins with a reasoning block wrapped in <think> tags. If you were planning to use it for cheap classification or extraction, budget for the reasoning tokens, because there is no way to turn them off on the open weights. The hosted service is the only place a non-thinking mode exists. LFM2.5-24B-A2B makes the same choice at the small end, with a chat template that starts every answer inside a <think> block. The same pattern shows up in closed models, where the Gemini Flash line dropped its lowest thinking level: within the Gemini 3.x Flash line, minimal still answers on 3.6 Flash while both 3.7 and 3.8 Flash reject it outright.
Which Model Should You Choose?
The “best” model depends entirely on what you’re building. Here’s a quick decision guide based on the verified numbers above and our own runs.
Best for coding: GLM-5.3 for terminal and agentic work, where it posts 88.2 on Terminal Bench 2.1 and 28.3 on Terminal Bench 3.0 against GLM-5.2’s 4.6, and DeepSeek V4 Pro if your yardstick is SWE-bench Verified, where its 80.6% still stands. On a single consumer GPU, Qwen3.8-27B is now the pick over Qwen3.6-27B, and it is still Apache 2.0, with Muse Glimmer 30B the Apache 2.0 alternative from Meta at 76.0% SWE-bench Verified. DeepSeek V4.1-Flash posts the best open-weight Terminal-Bench 2.1 figure on its own card, 90.6, if you can host 510 GB.
Best for reasoning and math: GLM-5.3 or DeepSeek V4.1-Flash, tied on 45 on the Artificial Analysis index, with Kimi K3 a point behind. DeepSeek R1 remains excellent for pure math on smaller budgets.
Best on a single 80 GB GPU: gpt-oss-120b at 65 GB was built for exactly this slot. Qwen3.8-27B now competes for it at 56 GB in BF16 or 31 GB in the FP8 build, and it accepts images and video, which gpt-oss does not. A 96 GB workstation card such as the RTX PRO 6000 also holds MiniCPM-V 4.7 at full BF16 (70 GB), which no 32 GB consumer card can.
Best for edge and laptops: gpt-oss-20b if you want reasoning in 16 GB; LFM2.5-24B-A2B for tool-calling agents on a 16 GB GPU at 14.5 GB in Q4_K_M; Gemma 4 E2B/E4B if you want image and audio input on-device; Phi-4 Mini for long context in under 4 GB of weights.
Best for multimodal: Qwen3.8-27B under Apache 2.0 if you want vision in something you can actually host, GLM-5.3-Flash if you want a frontier-class multimodal model under MIT, Gemma 4 for audio input on the 12B and E-series, or MiniCPM-V 4.7 for image and video at 3B active under Apache 2.0. Muse Spark Open is the only open flagship that takes text, images, video and audio together, if you can find the 700 GB of GPU memory its card asks for. Check this column before committing, because the flagships split on it: GLM-5.3, Qwen3.8-Max and gpt-oss are all text-only, and so is the small LFM2.5.
Best for long context: Llama 4 Scout’s 10M tokens is still unmatched on paper. For a practical 1M with current-generation quality, Kimi K3, DeepSeek V4 and V4.1-Flash, GLM-5.3, or Muse Spark Open, all of which take 1M natively.
Best for multilingual: Qwen 3.5/3.6 (201 languages) or Gemma 4 (140+ languages, Apache 2.0).
Best permissive license: Qwen3.8-27B, Qwen 3.6, Gemma 4, MiniCPM-V 4.7, Muse Glimmer 30B, gpt-oss (Apache 2.0), DeepSeek V4 and V4.1-Flash, GLM-5.3-Flash, Phi-4 (MIT). No usage caps, no royalties, no geographic restrictions. Note that this is now a per-model question rather than a per-vendor one, and not a size rule either: Qwen3.8-27B stays Apache 2.0 and GLM-5.3-Flash stays MIT, but Qwen3.8-Flash-Next carries a stricter resale gate than its own larger sibling, and the strictest of any otherwise-permissive release here. Command A and Command R+ are tighter still, though they are non-commercial outright rather than permissive-with-a-gate. Kimi K3, Muse Spark Open, LFM2.5 and the Llama family carry gates worth reading twice.
Running These Models Locally with Ollama
Every model in the March test table, plus gpt-oss and the Qwen 3.6 releases, can be pulled and run with a single Ollama command, and the library now lists LFM2.5-24B-A2B as lfm2.5:24b. Install Ollama first:
curl -fsSL https://ollama.com/install.sh | shThen pull and interact with any model:
ollama run gpt-oss:20bOn a GPU-equipped machine, Ollama automatically uses CUDA or ROCm, cutting response times from minutes to seconds. On CPU-only systems, stick with models under 8B parameters for usable response times, or step up to a unified-memory box for the 100 GB class; our Mac mini vs mini PC vs GPU comparison walks through that decision.
For a full setup walkthrough, see our guide on installing Ollama on Rocky Linux and Ubuntu. If you want a ChatGPT-style web interface for your local models, Open WebUI provides exactly that. And when Ollama’s abstractions get in the way of squeezing out performance, running llama.cpp directly buys you measurably higher throughput on the same hardware.
